Step 5 Preview vs Grok 4.5: 하나는 숫자를 보여주고, 다른 하나는 가중치를 보여준다

Grok 4.5 and S​tep 5 Preview are mirror images of each other, and the mirror is the reason to read this rather than the price. x​AI released Grok 4.5 on July 8, 2026; StepFun released S​tep 5 Preview on September 20, 2026. Grok 4.5 publishes a vendor benchmark you can quote — 64.7% on SWE-bench Pro — and an independent index score, and gives you no weights at all. S​tep 5 Preview publishes its weights in three precisions, ungated, and publishes no numeric benchmark table anywhere in its documentation, only a picture of a score chart on the model card. If you are choosing between them, the question is not which model is better. It is which kind of unknown you would rather carry: a model you can download but cannot score, or a model you can score but cannot download.

Everything else about the matchup resolves cleanly and mostly in StepFun's favour on paper. S​tep 5 Preview is half the price on both ends of the rate card, has twice the context window, streams faster on the one independent measurement both models have, and costs an order of magnitude less to keep a long prefix in cache. Grok 4.5's counterargument is not its own specs — it is that it is the middle of a family that has moved twice since July, and that its cache rate is the cheapest in that family.

두 요금표 모두, 모든 행

• Input price — S​tep 5 Preview $1.00 per million tokens vs Grok 4.5 $2.00 per million tokens

• Output price — S​tep 5 Preview $2.70 per million tokens vs Grok 4.5 $6.00 per million tokens

• Cache-hit input — S​tep 5 Preview $0.10 per million tokens, a 90% discount vs Grok 4.5 $0.30 per million tokens, an 85% discount, and the cheapest cache rate in the G​rok line

• Context window — S​tep 5 Preview 1M tokens vs Grok 4.5 500,000 tokens

• Maximum output — S​tep 5 Preview 64k tokens per StepFun's documentation, with the model card's own table claiming "up to 1M" vs Grok 4.5 not stated in the same terms on the vendor's pricing pages

• Long-prompt repricing — S​tep 5 Preview one flat rate across the whole 1M window vs Grok 4.5 doubles to $4.00 input and $12.00 output once a single request passes 200,000 tokens, applied to the whole request rather than the overage

• Architecture — S​tep 5 Preview sparse mixture-of-experts, 600B total and 27B active parameters across 352 experts with 8 selected per token vs Grok 4.5 a 1.5-trillion-parameter mixture-of-experts on the V9 foundation, roughly three times the size of Grok 4.3

• Input modality — S​tep 5 Preview text, images and video vs Grok 4.5 text, images and files

• Independent index — S​tep 5 Preview 44, at rank 41 of 219 when that page was captured vs Grok 4.5 39, on a page Artificial Analysis now marks deprecated, with the number frozen at the model's final reading

• Output speed — S​tep 5 Preview 88.1 output tokens per second vs Grok 4.5 60 tokens per second on the current reading, below the 88-token average and down from the 80 tokens per second x​AI quoted at launch

• Vendor benchmark — S​tep 5 Preview no numeric table published in the documentation, with the card's results presented as an image vs Grok 4.5 64.7% on SWE-bench Pro, vendor-reported

• Open weights — S​tep 5 Preview FP8, BF16 and NVFP4 checkpoints under StepFun's own account, ungated and downloadable since the morning of release vs Grok 4.5 none published

대칭성을, 정확히 서술하면

StepFun's documentation for S​tep 5 Preview contains an "At a glance" table with the model ID, the context window, the input and output types and the input and output limits. It does not contain a benchmark row. The evaluation section of the model card names the harnesses and the sampling settings — temperature 1.0, top-p 0.95, top-k 0, the High reasoning setting, SWE-agent for DeepSWE v1.1 — and then presents the results as a PNG. There is no number in StepFun's own documentation for this model that you can read as text and quote, which means every performance claim about S​tep 5 Preview in circulation traces back to either that image or to Artificial Analysis.

오픈 웨이트를 갖춘 모델이 놓이기에는 이상한 자리다. 오픈 웨이트는 보통 벤치마크 대신 얻는 것이지, 벤치마크 공백과 나란히 얻는 것이 아니기 때문이다. Artificial Analysis Intelligence Index에서의 44는 실제이고 독립적이며, 이 모델이 가진 유일한 중립 점수다. 그리고 평가 기관 자체 페이지에는 이 복합 지수 안에 포함된 여러 개별 평가에 '검토 중' 표시가 붙어 있다. 44를 프로필로 취급하지 말고, 출처가 탄탄한 단일 숫자로 취급하라.

Grok 4.5 is the exact inverse. Its weights are not published and never were, so the model is a black box you rent by the token. What you get in exchange is a documented 64.7% on SWE-bench Pro from the vendor, plus a live independent profile: 39 on the same index version, 60 output tokens per second on the current reading and therefore below average, $0.88 per completed index task, and 77 million output tokens across the suite against a median of 85 million. That last figure is the token-economy claim showing up in somebody else's measurement, and it holds — Grok 4.5 is the more concise model. It is also the reason its cost per task lands below S​tep 5 Preview's despite a rate card that is twice as expensive on both ends. One caveat that arrived with the model's age: the evaluator now flags the Grok 4.5 page as deprecated, keeps benchmarking only the default 10,000-token workload, and marks results for other workloads as historical rather than updated. The 39 is the model's final reading, not a live one.

그 둘은 모두 제대로 된 제품입니다. 다만 단일 축에서 비교할 수 있는 대상이 아니며, 두 벤더 중 하나는 그 수치를 자발적으로 공개하고 다른 하나는 공개하지 않는다는 점을 밝히지 않은 채 44를 39 옆에 놓는 어떤 점수판이든 더 중요한 차이를 감추고 있는 셈입니다. 44 대 39 격차는 서로 맞춰지지 않은 추론 노력에서 측정된 두 종합 점수 사이의 5점 차이일 뿐이고, 누구도 직접 맞대결로 실행한 적이 없습니다. 그것은 두 배를 지불할 이유가 아닙니다. 가중치는 선택할 이유가 되며, 가격도 마찬가지입니다.

왜 저렴한 모델이 빠른 것인가

Twenty-seven billion active parameters out of six hundred billion is what makes S​tep 5 Preview cheap to serve, and the serving economics show up in the independent measurement: 88.1 output tokens per second against Grok 4.5's current 60, on a model with twice the context window and half the price. Grok 4.5's 1.5-trillion-parameter V9 foundation is a third larger than Grok 4.3 and is the reason it has half of Grok 4.3's context at twice the input price — a bigger base model bought capability at the cost of the window, and by the time you are paying for it per token, that trade has a bill attached.

The latency picture is where Grok 4.5 has a defensible case and the independent numbers back it, which is worth stating rather than skipping. x​AI sold this model on token economy — fewer output tokens per finished task — and the measurement agrees: 77 million output tokens across the intelligence suite against a median of 85 million, with the evaluator describing it as "fairly concise." S​tep 5 Preview went the other way, generating 160 million tokens against the same 85-million median and drawing a "very verbose" label. A model that finishes a task in fewer tokens wins on wall-clock time even when it generates more slowly, and it wins on the invoice even when its rate card is worse. That is why Grok 4.5 costs $0.88 per completed index task at $2.00 and $6.00 per million, while S​tep 5 Preview pays for its cheaper rate card with verbosity. If your workload is a large number of short, well-specified calls, the token-economy argument is the reason to test Grok 4.5 on your own traffic rather than dismissing it on price.

캐시 비율은 Grok 4.5를 선택해야 하는 유일한 항목이다

Grok 4.5 costs three times what S​tep 5 Preview charges to read a cached token — thirty cents per million against ten — so this is not a G​rok advantage over StepFun. It is an advantage over G​rok's own successors. Grok 4.6 and Grok 4.7 both cache at $0.50 per million according to the vendor's pricing table, sixty-six percent more than Grok 4.5 charges, and all three list at the same $2.00 and $6.00 on input and output. That combination makes Grok 4.5 the only interesting member of its own family for one specific shape of workload: a large, stable, repeatedly-hit prefix on a pinned G​rok snapshot, where the cache line dominates the bill.

Against S​tep 5 Preview, even that line goes the other way. And the 200,000-token repricing threshold compounds it: a Grok 4.5 request at 250,000 tokens is billed at double rate on all 250,000 tokens, while the same request on S​tep 5 Preview is billed at $1.00 per million across the whole window. On a long-context pipeline, that is the difference between a flat rate and a cliff.

10월 체크포인트가 이 비교에 가져오는 변화

S​tep 5 Preview's weights have been downloadable since the morning of release — the FP8 repository was created at 03:15 UTC on September 20, 2026 and the BF16 repository at 03:52, with their shards committed the same morning — but the deployment story only became complete in mid-October. StepFun published an NVFP4 checkpoint on October 14, 2026, 394.18 GB configured for N​VIDIA's four-bit format with group_size: 16 and an unquantized KV cache, and the same week added a full serving guide to the older repositories: a published container image, worked tensor, expert and combined-parallelism commands, and dedicated reasoning and tool-call parsers so the model's own output format survives an O​penAI-shaped API. On October 15 N​VIDIA publicly called this "their largest open weights release" and reported that it "runs at nearly 8K tokens/sec/GPU on N​VIDIA GB300 using NVFP4" — a vendor figure on vendor hardware that nobody outside N​VIDIA has reproduced, attached to a repository you can inspect.

None of that makes S​tep 5 Preview the more measurable model; the benchmark gap stands. What it does is make the weights argument concrete. Grok 4.5 cannot be self-hosted, cannot be pinned to a version you control, and cannot be audited. S​tep 5 Preview can be all three, and the recipe is published. If a data-residency rule, a procurement rule or a model-deprecation risk sits anywhere in your decision, that is the row that decides it, and no amount of SWE-bench Pro closes it.

스택에서 각각이 어디에 속하는지

Grok 4.5 is on OrcaRouter as grok/grok-4.5, alongside the rest of the G​rok line and the rest of the catalogue, at the vendor's own $2.00 and $6.00 with the $0.30 cache rate passed through unchanged. That matters for the case Grok 4.5 is actually good at: a pinned snapshot with a long stable prefix, called alongside everything else on one OpenAI-compatible endpoint, one key and one bill, with 0% markup on the provider's list price so a price change reaches you the day it lands rather than on the next cycle. If you are already routing Grok 4.6 or Grok 4.7, comparing against 4.5 is a string change and a re-run of your own evals — automatic failover across providers means the pin is a routing rule rather than a dependency, and the DSL composes several models into one call when a step needs two opinions instead of a cheaper one.

S​tep 5 Preview is not on OrcaRouter's live model list, and the honest instruction is to call it on StepFun's own platform or self-host it from the published checkpoints. That is not a limitation of the router so much as a property of the model: the reason to pick it is the weights, and if you are using the weights you are not routing it anywhere.

두 제품이 같다고 가장하지 않은 결론

Take S​tep 5 Preview for the price, the context and the weights — $1.00 and $2.70 per million tokens, a flat rate across a million tokens, video input, faster generation on the only independent throughput measurement both models have, and three downloadable checkpoints with a published serving recipe. Accept that you will be buying it with less published performance evidence than you would like, and that the one number you have is a single independent composite with review flags on its components. It was released on September 20, 2026 and its deployment story got materially better in October; that is the model you would be adopting.

Take Grok 4.5 only if you specifically want a documented benchmark you can put in a procurement document, the most concise output profile of any model in this comparison, or the cheapest cache rate in the G​rok family on a pinned snapshot you already trust — and note that Grok 4.5 wins on cost per completed task despite losing on every line of the rate card, which is the strongest argument in this article for it. Understand what you are giving up: a two-generation-old model at the same list price as its two successors, no weights, a repricing cliff at 200,000 tokens, half the context of the cheaper alternative, and a score the evaluator has frozen. The upside is that you know what you are getting, and on a black-box model that is worth something. It is not worth paying double on input and output, but the task-level numbers say the premium is smaller than the rate card makes it look.

이 중 어느 것도 해당 모델을 피해야 할 이유는 아니며, 단지 그 이력을 알아야 할 이유일 뿐입니다: 이 계열에 속한 모델 뒤에 있는 문서화된 AI 사고와 에이전트 실패는 OrcaRouter AI Incident Archive에 수집되어 있으며, 각 항목은 자체 출처, 심각도, 신뢰도 등급 및 이의 제기 플래그를 포함합니다.